Also not sure yet how to determine the height for the weight to be changed at yet, but I can start at the center of gravity.
When you add weight to a side of the top, at the height of the center of gravity, you are shifting the vertical principal axis of inertia, maintaining it parallel to its previous position. You shift it without tilting it.
If you add weight to a side of the top, above or below the center of gravity, you are shifting
and tilting the vertical principal axis of inertia.
I had to think for a while if there is a difference by detecting the unbalance at the tip or at the stem, and I found that there is a difference in fact:
with your system, you are detecting the direction towards which you have to shift the vertical principal axis of inertia, for to make it pass through the tip. When the vertical principal axis of inertia passes through the tip, the top is balanced; the tip doesn't vibrate anymore, the spin times are increased, and the topple speed reduced.
But there is a third element, the top symmetry axis, which we would like to be superposed and coincident to the vertical principal axis of inertia, otherwise the top might look wobbling even if it is balanced.
Your system gives the informations for to align the vertical principal axis of inertia to the tip, but not those necessary for to align the vertical principal axis of inertia to the symmetry axis of the top.
Depending on how much finicky you are about it, this probably will be not a big problem anyway, you correct the unbalance just by adding, (or removing), weight at the height of the center of gravity and the result usually will be fine.
But if you have to balance a top which has the two axes tilted towards each other, you can't know in which direction the tilting of the vertical principal axis of inertia has to be corrected. In this sense, it would be better to detect the unbalance at the stem, because this allows to gather the informations for to align all the three elements and to make the top to be balanced, and, also, to look balanced.